Deyu Cao

I have worked on multiple projects and research initiatives in the field of machine learning, with a primary focus on generative models and the quantization of large language models (LLMs). My research interests center on model compression and energy-efficient deep learning from both algorithmic and hardware perspectives.
For my undergraduate thesis, I conducted research on optical neural networks, and from April 2026, I plan to pursue graduate studies at the University of Tokyo (EEIS), where I will focus on neuromorphic AI processor research.
I aim to deepen my understanding of both algorithms and hardware systems, and to contribute to driving the efficient and environmentally sustainable implementation of cutting-edge technologies.

• Outstanding Undergraduate Thesis Award, Department of Information and Communication Engineering, the University of Tokyo (Mar 2026)

Peer-reviewed conference papers/ journal papers

Daniel Saragih, Deyu Cao, Tejas Balaji, Ashwin Santhosh, "Flow to Learn: Flow Matching on Neural Network Parameters", Poster Presentation at ICLR Workshop on Neural Network Weights as a New Data Modality 2025.

Daniel Saragih, Deyu Cao, Tejas Balaji, "Generative Flow Models in Weight Space for Detecting Covariate Shifts",  (Planned) Oral Presentation at AAAI-26 Workshop.

Deyu Cao, Samin Aref, "Enhancing Ultra-Low-Bit Quantization of Large Language Models Through Saliency-Aware Partial Retraining", Oral Presentation at MDAI 2025.

Deyu Cao, Yixin Yin, Samin Aref, "Sliced-Wasserstein Distribution Alignment Loss Improves the

Ultra-Low-Bit Quantization of Large Language Models", Oral Presentation at ICAART 2026, Best Paper Award.

Domestic conference presentations

Design of Dielectric Metasurface-based Hybrid Optical Neural Networks for Image Generation, Photonic Device Workshop 2025, Best Student Poster Award.

Design and Fabrication of Dielectric Metasurface for Incoherent Digital Holography, OPE 2025, Outstanding Student Poster Award.